Incorporating Lexico-semantic Heuristics into Coreference Resolution Sieves for Named Entity Recognition at Document-level

Incorporating Lexico-semantic Heuristics into Coreference Resolution Sieves for Named Entity Recognition at Document-level Marcos Garcia Grupo LyS, Dep. de Galego-Portugu´es, Franc´es e Ling¨u´ıstica Fac. de Filolox´ıa, Universidade da Coru˜na Campus da Coru˜na, 15701, Coru˜na, Galicia, Spain marcos.garcia.gonzalez@udc.gal Abstract This paper explores the incorporation of lexico-semantic heuristics into a deterministic Coreference Resolution (CR) system for classifying named entities at document-level. The highest precise sieves of a CR tool are enriched with both a set of heuristics for merging named entities labeled with different classes and also with some constraints that avoid the incorrect merging of similar mentions. Several tests show that this strategy improves both NER labeling and CR. The CR tool can be applied in combination with any system for named entity recognition using the CoNLL format, and brings benefits to text analytics tasks such as Information Extraction. Experiments were carried out in Spanish, using three different NER tools. Keywords: named entity recognition, coreference resolution, information extraction 1. Introduction Most Named Entity Recognition (NER) systems label each instance of a Named Entity (NE) using only local informa- tion (i.e., from the immediate context of the analyzed to- ken). This strategy involves the misclassification of some mentions of the same entity in a single document.1 For instance, a common NER system may classify most men- tions of the entity “Lionel Messi” (“Messi”, “Leo Messi”, etc.) in the same document as a person, but some of them might be wrongly labeled as an organization or a loca- tion. These misclassifications also harm subsequent pro- cesses such as Coreference Resolution (CR) or Information Extraction (IE), which heavily depend on the accuracy of NER. Although some methods make use of global information (from the whole document) for NER (Finkel et al., 2005), these systems might also classify mentions of the same en- tity with different labels. Besides, many NERs depend on other NLP modules such as tokenizers, lemmatizers or PoS- taggers, whose adaptation to global information NERs re- quire a big effort. On the other hand, most common strategies for nominal CR include several matching heuristics which cluster dif- ferent mentions sharing —among other information— parts of their surface form: e.g., mentions such as “Lennon PER” and “John Lennon PER” might be merged into the same en- tity (Recasens and Hovy, 2009; Lee et al., 2013). How- ever, as these systems depend on the previous NER tools, it is not likely that they merge mentions belonging to dif- ferent classes, such as “John Winston Lennon PER” and “Lennon ORG” (wrongly classified). In order to improve both NER and CR, this paper explores the incorporation of lexico-semantic heuristics into the CR sieves with the highest precision. On one hand, this allows the CR tool to merge nominal mentions (which actually be- 1Following Recasens and Mart´ı (2010), a mention is each in- stance of reference to an object, while an entity is the collection of mentions referring to the same object in a document. long to the same entity) which had been misclassified with different NER …

Content truncated to 500 words